The Eect of Advertising on Brand Awareness and Perceived Quality: An Empirical Investigation using Panel Data

C. Robert Clark Ulrich Doraszelski February 21, 2009 Michaela Draganska

Abstract We use a panel data set that combines annual brand-level advertising expenditures for over three hundred brands with measures of brand awareness and perceived quality from a large-scale consumer survey to study the eect of advertising. Advertising is modeled as a dynamic investment in a brands stocks of awareness and perceived quality and we ask how such an investment changes brand awareness and quality perceptions. Our panel data allow us to control for unobserved heterogeneity across brands and to identify the eect of advertising from the time-series variation within brands. They also allow us to account for the endogeneity of advertising through recently developed dynamic panel data estimation techniques. We nd that advertising has consistently a signicant positive eect on brand awareness but no signicant eect on perceived quality.

The authors would like to thank Jason Allen, Christina Gathmann, Wes Hartmann, Ig Horstmann, Jordi Jaumandreu, Phillip Leslie, Puneet Manchanda, Julie Mortimer, Harikesh Nair, Peter Rossi, and two anonymous referees for helpful comments and Harris Interactive for providing the Equitrend data used in this study. Clark would like to thank the CIRPEE and FQRSC for research support for this project and Doraszelski gratefully acknowledges the hospitality of the Hoover Institution during the academic year 2006/07. Institute of Applied Economics, HEC Montreal and CIRPEE, 3000 Chemin de la Cote-SainteCatherine, Montreal (Quebec) H3T 2A7, robert.clark@hec.ca Department of Economics, Harvard University, 1805 Cambridge Street, Cambridge, MA 02138, doraszelski@harvard.edu Stanford University, Graduate School of Business, Stanford, CA 94305-5015, draganska michaela@gsb.stanford.edu

Introduction

In 2006 more than $280 billion were spent on advertising in the U.S., well above 2% of GDP. By investing in advertising, marketers aim to encourage consumers to choose their brand. For a consumer to choose a brand, two conditions must be satised: First, the brand must be in her choice set. Second, the brand must be preferred over all the other brands in her choice set. Advertising may facilitate one or both of these conditions. In this research we empirically investigate how advertising aects these two conditions. To disentangle the impact on choice set from that on preferences, we use actual measures of the level of information possessed by consumers about a large number of brands and of their quality perceptions. We compile a panel data set that combines annual brand-level advertising expenditures with data from a large-scale consumer survey, in which respondents were asked to indicate whether they were aware of dierent brands and, if so, to rate them in terms of quality. These data oer the unique opportunity to study the role of advertising for a wide range of brands across a number of dierent product categories. The awareness score measures how well consumers are informed about the existence and the availability of a brand and hence captures directly the extent to which the brand is part of consumers choice sets. The quality rating measures the degree of subjective vertical product dierentiation in the sense that consumers are led to perceive the advertised brand as being better. Hence, our data allow us to investigate the relationship between advertising and two important dimensions of consumer knowledge. The behavioral literature in marketing has highlighted the same two dimensions in the form of the size of the consideration set and the relative strength of preferences (Nedungadi 1990, Mitra & Lynch 1995). It is, of course, possible that advertising also aects other aspects of consumer knowledge. For example, advertising may generate some form of subjective horizontal product dierentiation that is unlikely to be reected in either brand awareness or perceived quality. In a recent paper Erdem, Keane & Sun (2008), however, report that advertising focuses on horizontal attributes only for one out of the 19 brands examined. Understanding the channel through which advertising aects consumer choice is important for researchers and practitioners alike for several reasons. For example, Suttons (1991) bounds on industry concentration in large markets implicitly assume that advertising increases consumers willingness to pay by altering quality perceptions. While prots increase in perceived quality, they may decrease in brand awareness

(Fershtman & Muller 1993, Boyer & Moreaux 1999), thereby stalling the competitive escalation in advertising at the heart of the endogenous sunk cost theory. Moreover, Doraszelski & Markovich (2007) show that even in small markets industry dynamics can be very dierent depending on the nature of advertising. From an empirical perspective, when estimating a demand model, advertising could be modeled as aecting the choice set or as aecting the utility that the consumer derives from a brand. If the role of advertising is mistakenly specied as aecting quality perceptions (i.e., preferences) rather than brand awareness as it often is, then the estimated parameters may be biased. In her study of the U.S. personal computer industry, Sovinsky Goeree (2008) nds that traditional demand models overstate price elasticities because they assume that consumers are aware of and hence choose among all brands in the market when in actuality most consumers are aware of only a small fraction of brands. For our empirical analysis we develop a dynamic estimation framework. Brand awareness and perceived quality are naturally viewed as stocks that are built up over time in response to advertising (Nerlove & Arrow 1962). At the same time, these stocks depreciate as consumers forget past advertising campaigns or as an old campaign is superseded by a new campaign. Advertising can thus be thought of as an investment in brand awareness and perceived quality. The dynamic nature of advertising leads us to a dynamic panel data model. In estimating this model we confront two important problems, namely unobserved heterogeneity across brands and the potential endogeneity of advertising. We discuss these below. When estimating the eect of advertising across brands we need to keep in mind that they are dierent in many respects. Unobserved factors that aect both advertising expenditures and the stocks of perceived quality and awareness may lead to spurious positive estimates of the eect of advertising. Put dierently, if we detect an eect of advertising, then we cannot be sure if this eect is causal in the sense that higher advertising expenditures lead to higher brand awareness and perceived quality or if it is spurious in the sense that dierent brands have dierent stocks of perceived quality and awareness as well as advertising expenditures. For example, although in our data the brands in the fast food category on average have high advertising and high awareness and the brands in the cosmetics and fragrances category have low advertising and low awareness, we cannot infer that advertising boosts awareness. We can only conclude that the relationship between advertising expenditures, perceived quality, and brand awareness diers from category to category or even from brand to brand. Much of the existing literature uses cross-sectional data to discern a relationship 3

between advertising expenditures and perceived quality (e.g., Kirmani & Wright 1989, Kirmani 1990, Moorthy & Zhao 2000, Moorthy & Hawkins 2005) in an attempt to test the idea that consumers draw inferences about the brands quality from the amount that is spent on advertising it (Nelson 1974, Milgrom & Roberts 1986, Tellis & Fornell 1988). With cross-sectional data it is dicult to account for unobserved heterogeneity across brands. Indeed, if we neglect permanent dierences between brands, then we nd that both brand awareness and perceived quality are positively correlated with advertising expenditures, thereby replicating the earlier studies. Once we make full use of our panel data and account for unobserved heterogeneity, however, the eect of advertising expenditures on perceived quality disappears.1 Our estimation equations are dynamic relationships between a brands current stocks of perceived quality and awareness on the left-hand side and the brands previous stocks of perceived quality and awareness as well as its own and its rivals advertising expenditures on the right-hand side. In this context, endogeneity arises for two reasons. First, the lagged dependent variables are by construction correlated with all past error terms and therefore endogenous. As a consequence, traditional xed-eect methods are necessarily inconsistent.2 Second, advertising expenditures may also be endogenous for economic reasons. For instance, media coverage such as news reports may aect brand awareness and perceived quality beyond the amount spent on advertising. To the extent that these shocks to the stocks of perceived quality and awareness of a brand feed back into decisions about advertising, say because the brand manager opts to advertise less if a news report has generated sucient awareness, they give rise to an endogeneity problem. To resolve the endogeneity problem we use the dynamic panel data methods developed by Arellano & Bond (1991), Arellano & Bover (1995), and Blundell & Bond (1998). The key advantage is that these methods do not rely on the availability of strictly exogenous explanatory variables or instruments. This is an appealing methodology that has been widely applied (e.g., Acemoglu & Robinson 2001, Durlauf, Johnson & Temple 2005, Zhang & Li 2007) because valid instruments are often hard to come by. Further, since these methods involve rst dierencing, they allow us to control for unobserved factors that aect both advertising expenditures and the stocks
Another way to get around this issue is to take an experimental approach, as in Mitra & Lynch (1995). 2 This source of endogeneity is not tied to advertising in particular; rather it always arises in estimating dynamic relationships in the presence of unobserved heterogeneity. An exception is the (rather unusual) panel-data setting where one has T instead of N . In this case the within-groups estimator is consistent (Bond 2002, p. 5).
1

of perceived quality and awareness and may lead to spurious positive estimates of the eect of advertising. In addition, our approach allows for factors other than advertising to aect a brands stock of perceived quality and awareness to the extent that these factors are constant over time. Our main nding is that advertising expenditures have a signicant positive effect on brand awareness but no signicant eect on perceived quality. These results appear to be robust across a wide range of specications. Since awareness is the most basic kind of information a consumer can have for a brand, we conclude that an important role of advertising is information provision. On the other hand, our results indicate that advertising is not likely to alter consumers quality perceptions. This conclusion calls for a reexamination of the implicit assumption underlying Suttons (1991) endogenous sunk cost theory. It also suggests that advertising should be modeled as aecting the choice set and not just utility when estimating demand. Finally, our ndings lend empirical support to the view that advertising is generally procompetitive because it disseminates information about the existence, the price, and the attributes of products more widely among consumers (Stigler 1961, Telser 1964, Nelson 1970, Nelson 1974). The remainder of the paper proceeds as follows. In Sections 2 and 3 we explain the dynamic investment model and the corresponding empirical strategy. In Section 4 we describe the data and in Section 5 we present the results of the empirical analysis. Section 6 concludes.

Model Specication

We develop an empirical model based on the classic advertising-as-investment model of Nerlove & Arrow (1962). Related empirical models are the basis of current research on advertising (e.g., Naik, Mantrala & Sawyer 1998, Dube, Hitsch & Manchanda 2005, Doganoglu & Klapper 2006, Bass, Bruce, Majumdar & Murthi 2007). Naik et al. (1998), in particular, nd that the Nerlove & Arrow (1962) model provides a better t than other models that have been proposed in the literature such as Vidale & Wolfe (1957), Brandaid (Little 1975), Tracker (Blattberg & Golanty 1978), and Litmus (Blackburn & Clancy 1982). We extend the Nerlove & Arrow (1962) framework in two respects. First, we allow a brands stocks of awareness and perceived quality to be aected by the advertising of its competitors. This approach captures the idea that advertising takes place in a competitive environment where brands vie for the attention of consumers. The 5

advertising of competitors may also be benecial to a brand if it draws attention to the entire category and thus expands the relevant market for the brand (e.g., Nedungadi 1990, Kadiyali 1996). Second, we allow for a stochastic component in the eect of advertising on the stocks of awareness and perceived quality to reect the success or failure of an advertising campaign and other unobserved inuences such as the creative quality of the advertising copy, media selection, or scheduling. More formally, we let Qit be the stock of perceived quality of brand i at the start of period t and Ait the stock of its awareness. We further let Eit1 denote the advertising expenditures of brand i over the course of period t 1 and Eit1 = (E1t1 , . . . , Ei1t1 , Ei+1t1 , . . . , Ent1 ) the advertising expenditures of its competitors. Then, at the most general level, the stocks of perceived quality and awareness of brand i evolve over time according to the laws of motion Qit = git (Qit1 , Eit1 , Eit1 , it ), Ait = hit (Ait1 , Eit1 , Eit1 , it ), where git () and hit () are brand- and time-specic functions. The idiosyncratic error it captures the success or failure of an advertising campaign along with all other omitted factors. For example, the quality of the advertising campaign may matter just as much as the amount spent on it. By recursively substituting for the lagged stocks of perceived quality and awareness we can write the current stocks as functions of all past advertising expenditures and the current and all past error terms. This shows that these shocks to brand awareness and perceived quality are persistent over time. For example, the eect of a particularly good (or bad) advertising campaign may linger and be felt for some time to come. We model the eect of competitors advertising on brand awareness and perceived quality in two ways. First, we consider a brands share of voice. We use its advertising expenditures, Eit1 , relative to the average amount spent on advertising by rival brands in the brands subcategory or category, E it1 .3 To the extent that brands compete with each other for the attention of consumers, a brand may have to outspend its rivals to cut through the clutter. If so, then what is important may not be the absolute amount spent on advertising but the amount relative to rival brands. Second, we consider the amount of advertising in the entire market
The Brandweek Superbrands survey reports on only the top brands (in terms of sales) in each subcategory or category. The number of brands varies from 3 for some subcategories to 10 for others. We therefore use the average, rather than the sum, of competitors advertising.
3

by including the average amount spent on advertising by rival brands in the brands subcategory or category. Advertising is market expanding if it attracts consumers to the entire category but not necessarily to a particular brand. In this way, competitors advertising may have a positive inuence on, say, brand awareness. Taken together, our estimation equations are Qit = i + t + Qit1 + f (Eit1 , E it1 ) + it , Ait = i + t + Ait1 + f (Eit1 , E it1 ) + it . (1) (2)

Here i is a brand eect that captures unobserved heterogeneity across brands and t is a time eect to control for possible systematic changes over time. The time eect may capture, for example, that consumers are systematically informed about a larger number of brands due to the advent of the internet and other alternative media channels. Through the brand eect we allow for factors other than advertising to aect a brands stocks of perceived quality and awareness to the extent that these factors are constant over time. For example, consumers may hear about a brand and their quality perceptions may be aected by word of mouth. Similarly, it may well be the case that consumers in the process of purchasing a brand become more informed about it and that their quality perceptions change, especially for high-involvement brands. Prior to purchasing a car, say, many consumers engage in research about the set of available cars and their respective characteristics, including quality ratings from sources such as car magazines and Consumer Reports. If these eects do not vary over time, then we fully account for them in our estimation because the dynamic panel data methods we employ involve rst dierencing. The parameter measures how much of last periods stocks of perceived quality and awareness are carried forward into this periods stocks; 1 can therefore be interpreted as the rate of depreciation of these stocks. Note that in the estimation we allow all parameters to be dierent across our estimation equations. For example, we do not presume that the carryover rates for perceived quality and brand awareness are the same. The function f () represents the response of brand awareness and perceived quality to the advertising expenditures of the brand and potentially also those of its rivals. In the simplest case absent competition we specify this function as
2 f (Eit1 ) = 1 Eit1 + 2 Eit1 .

This functional form is exible in that it allows for a nonlinear eect of advertising expenditures but does not impose one. Later on in Section 5.6 we demonstrate the robustness of our results by considering a number of additional functional forms. To account for competition in the share-of-voice specication, we set f (Eit1 , E it1 ) = 1 Eit1 E it1 + 2 Eit1 E it1
2

and in the total-advertising specication, we set

2 f (Eit1 , E it1 ) = 1 Eit1 + 2 Eit1 + 3 E it1 .

Estimation Strategy

Equations (1) and (2) are dynamic relationships that feature lagged dependent variables on the right-hand side. When estimating, we confront the problems of unobserved heterogeneity across brands and the endogeneity of advertising. In our panel-data setting, ignoring unobserved heterogeneity is akin to dropping the brand eect i from equations (1) and (2) and then estimating them by ordinary least squares. Since this approach relies on both cross-sectional and time-series variation to identify the eect of advertising, we refer to it as pooled OLS(POLS) in what follows. To account for unobserved heterogeneity we include a brand eect i and use the within estimator that treats i as a xed eect. We follow the usual convention in microeconomic applications that the term xed eect does not necessarily mean that the eect is being treated as nonrandom; rather it means that we are allowing for arbitrary correlation between the unobserved brand eect and the observed explanatory variables (Wooldridge 2002, p. 251). The within estimator eliminates the brand eect by subtracting the within-brand mean from equations (1) and (2). Hence, the identication of the slope parameters that determine the eect of advertising relies solely on variation over time within brands; the information in the between-brand cross-sectional relationship is not used. We refer to this approach as xed eects (FE). While FE accounts for unobserved heterogeneity, it suers from an endogeneity problem. In our panel-data setting, endogeneity arises for two reasons. First, since equations (1) and (2) are inherently dynamic, the lagged stocks of perceived quality and awareness may be endogenous. More formally, Qit1 and Ait1 are by construc8

tion correlated with is for s < t. The within estimator subtracts the within-brand mean from equations (1) and (2). The resulting regressor, say Qit1 Qi in the case of perceived quality, is correlated with the error term it i since i contains it1 along with all higher-order lags. Hence, FE is necessarily inconsistent. Second, advertising expenditures may also be endogenous for economic reasons. For instance, media coverage such as news reports may directly aect brand awareness and perceived quality. Our model treats media coverage other than advertising as shocks to the stocks of perceived quality and awareness. To the extent that these shocks feed back into decisions about advertising, say because the brand manager opts to advertise less if a news report has generated sucient awareness, they give rise to an endogeneity problem. More formally, it is reasonable to assume that Eit1 , the advertising expenditures of brand i over the course of period t 1, are chosen at the beginning of period t 1 with knowledge of it1 and higher-order lags and that therefore Eit1 is correlated with is for s < t. We apply the dynamic panel-data method proposed by Arellano & Bond (1991) to deal with both unobserved heterogeneity and endogeneity. This methodology has the advantage that it does not rely on the availability of strictly exogenous explanatory variables or instruments. This is welcome because instruments are often hard to come by, especially in panel-data settings: The problem is nding a variable that is a good predictor of advertising expenditures and is uncorrelated with shocks to brand awareness and perceived quality; nding a variable that is a good predictor of lagged brand awareness and perceived quality and uncorrelated with current shocks to brand awareness and perceived quality is even less obvious. The key idea of Arellano & Bond (1991) is that if the error terms are serially uncorrelated, then lagged values of the dependent variable and lagged values of the endogenous right-hand-side variables represent valid instruments. To see this, take rst dierences of equation (1) to obtain Qit Qit1 = (t t1 ) + (Qit1 Qit2 ) + (f (Eit1 ) f (Eit2 )) + (it it1 ), (3) where we abstract from competition to simplify the notation. Eliminating the brand eect i accounts for unobserved heterogeneity between brands. The remaining problem with estimating equation (3) by least-squares is that Qit1 Qit2 is by construction correlated with it it1 since Qit1 is correlated with it1 by virtue of equation (1). Moreover, as we have discussed above, Eit1 may also be correlated with it1 for economic reasons.

We take advantage of the fact that we have observations on a number of periods in order to come up with instruments for the endogenous variables. In particular, this is possible starting in the third period where equation (3) becomes Qi3 Qi2 = (3 2 ) + (Qi2 Qi1 ) + (f (Ei2 ) f (Ei1 )) + (i3 i2 ). In this case Qi1 is a valid instrument for (Qi2 Qi1 ) since it is correlated with (Qi2 Qi1 ) but uncorrelated with (i3 i2 ) and, similarly, Ei1 is a valid instrument for (f (Ei2 ) f (Ei1 )). In the fourth period Qi1 and Qi2 are both valid instruments since neither is correlated with (i4 i3 ) and, similarly, Ei1 and Ei2 are both valid instruments. In general, for lagged dependent variables and for endogenous righthand-side variables, levels of these variables that are lagged two or more periods are valid instruments. This allows us to generate more instruments for later periods. The resulting estimator is referred to as dierence GMM (DGMM). A potential diculty with the DGMM estimator is that lagged levels may be poor instruments for rst dierences when the underlying variables are highly persistent over time. Arellano & Bover (1995) and Blundell & Bond (1998) propose an augmented estimator in which the original equations in levels are added to the system. The idea is to create a stacked data set containing dierences and levels and then to instrument dierences with levels and levels with dierences. The required assumption is that brand eects are uncorrelated with changes in advertising expenditures. This estimator is commonly referred to as system GMM (SGMM). In Section 5 we report and compare results for DGMM and SGMM. It is important to test the validity of the instruments proposed above. Following Arellano & Bond (1991) we report a Hansen J test for overidentifying restrictions. This test examines whether the instruments are jointly exogenous. We also report the so-called dierence-in-Hansen J test to examine specically whether the additional instruments for the level equations used in SGMM (but not in DGMM) are valid. Arellano & Bond (1991) further develop a test for second-order serial correlation in the rst dierences of the error terms. As described above, both GMM estimators require that the levels of the error terms be serially uncorrelated, implying that the rst dierences are serially correlated of at most rst order. We caution the reader that the test for second-order serial correlation is formally only dened if the number of periods in the sample is greater than or equal to 5 whereas we observe a brand on average for just 4.2 periods in our application. Our preliminary estimates suggest that the error terms are unlikely to be serially

10

uncorrelated as required by Arellano & Bond (1991). The AR(2) test described above indicates rst-order serial correlation in the error terms. An AR(3) test for thirdorder serial correlation in the rst dierences of the error terms, however, indicates the absence of second-order serial correlation in the error terms.4 In this case, Qit2 and Eit2 are no longer valid instruments for equation (3). Intuitively, because Qit2 is correlated with it2 by virtue of equation (1) and it2 is correlated with it1 by rst-order serial correlation, Qit2 is correlated with it1 in equation (3), and similarly for Eit2 . Fortunately, however, Qit3 and Eit3 remain valid instruments because it3 is uncorrelated with it1 . We carry out the DGMM and SGMM estimation using STATAs xtabond2 routine (Roodman 2007). We enter third and higher lags of either brand awareness or perceived quality, together with third and higher lags of advertising expenditures as instruments. In addition to these GMM-style instruments, for the dierence equations we enter the time dummies as IV-style instruments. We also apply the nite-sample correction proposed by Windmeijer (2005) which corrects for the two-step covariance matrix and substantially increases the eciency of both GMM estimators. Finally, we compute standard errors that are robust to heteroskedasticity and arbitrary patterns of serial correlation within brands.

Data

Our data are derived from the Brandweek Superbrands surveys from 2000 to 2005. Each years survey lists the top brands in terms of sales during the past year from 25 broad categories. Inside these categories are often a number of more narrowly dened subcategories. Table 1 lists the categories along with their subcategories. The surveys report perceived quality and awareness scores for the current year and the advertising expenditures for the previous year by brand. Perceived quality and awareness scores are calculated by Harris Interactive in their Equitrend brand-equity study. Each year Harris Interactive surveys online between 20, 000 and 45, 000 consumers aged 15 years and older in order to determine their perceptions of a brands quality and its level of awareness for approximately 1, 000 brands.5 To ensure that the respondents accurately reect the general popOf course, the AR(3) test uses less observations than the AR(2) test and is therefore also less powerful. 5 The exact wording of the question is: We will display for you a list of brands and we are asking you to rate the overall quality of each brand using a 0 to 10 scale, where 0 means Unacceptable/Poor Quality, 5 means Quite Acceptable Quality and 10 means Outstanding/Extraordinary Quality.
4

11

1. Apparel 2. Appliances 3. Automobiles a. general automobiles b. luxury c. subcompact d. sedan/wagon e. trucks/suvs/vans 4. Beer, Wine, Liquor a. beer b. wine c. malternatives d. liquor 5. Beverages a. general b. new age/sports/water 6. Computers a. software b. hardware 7. Consumer Electronics 8. Cosmetics and Fragrances a. color cosmetics b. eye color c. lip color d. womens fragrances e. mens fragrances 9. Credit Cards 10. Entertainment 11. Fast Food 12. Financial Services 13. Food a. ready to eat cereal b. cereal bars c. cookies d. cheese e. crackers f. salted snacks g. frozen dinners and entres e

14. 15.

16.

17.

18.

19. 20. 21. 22. 23. 24. 25.

h. frozen pizza i. spaghetti sauce j. coee k. ice cream l. refrigerated orange juice m. refrigerated yogurt n. soy drinks o. luncheon meats p. meat alternatives q. baby formula/electrolyte solutions r. pourable salad dressing Footwear Health and Beauty a. bar soap b. toothpaste c. shampoo d. hair color Household a. cleaner b. laundry detergents c. diapers d. facial tissue e. toilet tissue f. automatic dishwater detergent Petrol a. oil companies b. automotive aftercare/lube Pharmaceutical OTC a. allergy/cold medicine b. stomach/antacids c. analgesics Pharmaceutical Prescription Retail Telecommunications Tobacco Toys Travel World Wide Web

Table 1: Categories and subcategories. Items in italics have been removed.

12

ulation their responses are propensity weighted. Each respondent rates around 80 of these brands. Perceived quality is measured on a 0-10 scale, with 0 meaning unacceptable/poor and 10 meaning outstanding/extraordinary. Awareness scores vary between 0 and 100 and equal the percentage of respondents that can rate the brands quality. The quality rating is therefore conditional on the respondent being aware of the brand.6 We supplement the awareness and quality measures with advertising expenditures that are taken from TNS Media Intelligence and Competitive Media Reporting. These advertising expenditures encompass spending in a wide range of media: Magazines (consumer magazines, Sunday magazines, local magazines, and business-to-business magazines), newspaper (local and national newspapers), television (network TV, spot TV, syndicated TV, and network cable TV), radio (network, national spot, and local), Spanish-language media (magazines, newspapers, and TV networks), internet, and outdoor. After eliminating categories and subcategories where observations are not at the brand level (apparel, entertainment, nancial services, retail, world wide web) or where the data are suspect (tobacco), we are left with 19 categories (see again Table 1). We then drop all private labels and all brands for which we do not have perceived quality and awareness scores as well as advertising expenditures for at least two years running. This leaves us with 348 brands. Table 2 contains descriptive statistics for the overall sample and also by category. In the overall sample the average awareness score is 69.35 and the average perceived quality score is 6.36. The average amount spent on advertising is around $66 million per year. There is substantial variation in these measures across categories. The variation in perceived quality (coecient of variation is 0.11 overall, ranging from 0.04 for appliances to 0.13 for computers) tends to be lower than the variation in brand awareness (coecient of variation is 0.28 overall, ranging from 0.05 for appliances to 0.46 for telecommunications), in line with the fact the quality rating is conditional on the respondent being aware of the brand. The contemporaneous correlation between
You may use any number from 0 to 10 to rate the brands, or use 99 for No Opinion option if you have absolutely no opinion about the brand. Panelists are being incentivized through sweepstakes on a periodic basis but are not paid for a particular survey. 6 The 2000 Superbrands survey does not separately report perceived quality and salience scores. We received these scores directly from Harris Interactive. 2000 is the rst year for which we have been able to obtain perceived quality and salience scores for a large number of brands. Starting with the 2004 and 2005 Superbrands surveys, salience is replaced by a new measure called familiarity. For these two years we received salience scores directly from Harris Interactive. The contemporaneous correlation between salience and familiarity is 0.98 and signicant with a p-value of 0.000.

13

brand awareness and perceived quality is 0.60 and signicant with a p-value of 0.000. The contemporaneous correlation between advertising expenditures and the change in brand awareness is 0.0488 and signicant with a p-value of 0.0985 and the contemporaneous correlation between advertising expenditures and the change in perceived quality is 0.0718 and signicant with a p-value of 0.0150. These correlations anticipate the spurious correlation between both brand awareness and perceived quality and advertising expenditures if permanent dierences between brands are neglected (POLS estimator). We will see though that the eect of advertising expenditures on perceived quality disappears once unobserved heterogeneity is accounted for (FE and GMM estimators). The intertemporal correlation is 0.98 for brand awareness, 0.95 for perceived quality, and 0.93 for advertising expenditures. This limited amount of intertemporal variation warrants preferring the SGMM over the DGMM estimator. At the same time, however, it constrains how nely we can slice the data, e.g., by isolating a brand-specic eect of advertising expenditures on brand awareness and perceived quality. Since the FE, DGMM, and SGMM estimators rely on within-brand across-time variation, it is important to ensure that there is a sucient amount of within-brand variation in brand awareness, perceived quality, and advertising expenditures. Table 3 presents a decomposition of the standard deviation in these variables into an acrossbrands and a within-brand component for the overall sample and also by category. The across-brands standard deviation is a measure of the cross-sectional variation and the within-brand standard deviation is a measure of the time-series variation. The across-brands standard deviation of brand awareness is about 6 times larger than the within-brand standard deviation. This ratio varies across categories and ranges from 2 for automobiles, beer, wine, liquor, and pharmaceutical prescription to 6 for health and beauty and pharmaceutical OTC. In case of perceived quality the ratio is about 4 (ranging from 1 for telecommunications to 5 for consumer electronics, credit cards, and household). Hence, while there is more cross-sectional than time-series variation in our sample, the time-series variation is substantial for both brand awareness and perceived quality. Figure 1 illustrates the decomposition for the overall sample. The left panels show histograms of the brand-mean of brand awareness, perceived quality, and advertising expenditures and the right panels show histograms of the de-meaned variables. Again it is evident that the time-series variation is substantial for both brand awareness and perceived quality.

14

15 Table 2: Descriptive statistics.

overall Appliances Automobiles Beer, Wine, Liquor Beverages Computers Consumer Electronics Cosmetics and Fragrances Credit Cards Fast Food Food Footwear Health and Beauty Household Petrol Pharmaceutical OTC Pharmaceutical Prescription Telecommunications Toys Travel

# obs # brands 1478 348 21 4 137 30 98 24 95 22 79 17 29 7 70 19 29 6 60 12 247 65 38 8 54 11 128 31 48 13 56 15 31 10 52 11 25 5 181 38

brand perceived awareness (0-100) quality (0-10) mean std. dev. mean std. dev. 69.35 19.43 6.36 0.70 85.09 4.54 7.35 0.32 67.81 6.72 6.51 0.59 62.23 10.13 5.68 0.72 84.57 13.84 6.51 0.58 59.80 23.05 6.41 0.81 67.83 18.68 6.60 0.73 49.37 15.75 5.83 0.52 70.97 18.08 6.24 0.73 93.83 5.32 6.28 0.42 80.18 14.94 6.66 0.65 64.95 18.98 6.39 0.42 82.50 9.80 6.67 0.41 73.83 16.03 6.66 0.56 60.52 17.19 5.95 0.30 76.96 13.89 6.79 0.37 29.97 9.69 5.54 0.67 49.33 22.86 5.28 0.52 72.12 9.74 6.95 0.32 59.48 15.43 6.26 0.52

advertising ($1,000,000) mean std. dev. 66.21 118.52 41.87 33.19 99.85 64.62 36.78 45.11 41.33 42.19 130.43 130.07 104.83 160.66 38.02 47.48 174.54 109.77 214.80 156.23 13.93 13.81 40.27 46.89 27.28 33.44 21.80 25.43 33.54 34.65 38.71 18.13 76.23 36.40 367.93 360.54 108.55 54.36 25.41 25.88

16 Table 3: Variance decomposition.

brand awareness (0-100) across within overall 20.117 3.415 Appliances 5.282 1.334 Automobiles 6.209 3.281 Beer, Wine, Liquor 10.181 4.105 Beverages 13.435 2.915 Computers 23.094 3.843 Consumer Electronics 19.952 5.611 Cosmetics and Fragrances 18.054 3.684 Credit Cards 19.568 3.903 Fast Food 6.132 1.660 Food 16.241 2.255 Footwear 20.417 4.267 Health and Beauty 10.536 1.772 Household 16.719 3.896 Petrol 20.179 3.669 Pharmaceutical OTC 13.339 2.363 Pharmaceutical Prescription 9.393 5.772 Telecommunications 21.659 5.604 Toys 11.217 3.589 Travel 16.063 3.216

perceived quality (0-10) across within 0.726 0.176 0.323 0.148 0.561 0.141 0.705 0.186 0.582 0.190 0.850 0.313 0.800 0.167 0.563 0.208 0.788 0.159 0.361 0.202 0.702 0.134 0.388 0.167 0.397 0.136 0.561 0.113 0.415 0.116 0.336 0.129 0.753 0.230 0.452 0.334 0.360 0.127 0.516 0.153

advertising ($1,000,000) across within 100.823 43.625 28.965 21.316 54.680 32.552 41.713 12.406 37.505 13.372 110.362 65.909 105.249 114.381 38.446 20.053 118.059 43.415 159.306 33.527 15.655 7.998 45.791 7.640 27.054 19.075 18.789 16.672 27.227 20.496 16.325 9.080 38.648 27.919 317.434 178.406 61.419 18.584 22.136 10.909

.02

.025

Density .015

.01

.005

20

40 60 Mean brand awareness

80

100

0 30

.05

Density .1

.15

.2

20

10 0 10 Demeaned brand awareness

20

30

.8

.6

Density .4

Density 0 2 4 6 Mean perceived quality 8 10 0 1.5 1

.2

.5 0 .5 Demeaned perceived quality

1.5

.015

.01

.005

200

400 600 800 1000 1200 Mean advertising expenditures (millions of $)

1400

0 600

.005

Density .01 .015

Density

.02

.025

400 200 0 200 400 Demeaned advertising expenditures (millions of $)

600

Figure 1: Variance decomposition. Histogram of brand-mean of brand awareness, perceived quality, and advertising expenditures (left panels) and histogram of de-meaned brand awareness, perceived quality, and advertising expenditures (right panels).

Empirical Results

In Tables 4 and 5 we present a number of dierent estimates for the eect of advertising expenditures on brand awareness and perceived quality, respectively. Starting with the simplest case absent competition, we present estimates of , 1 , and 2 (the 2 coecients on Qit1 or Ait1 and Eit1 and Eit1 ) along with the marginal eect 1 + 22 Eit1 calculated at the mean and the 25th, 50th, and 75th percentiles of advertising expenditures. The POLS estimates in the rst column of Tables 4 and 5 suggest a signicant positive eect of advertising expenditures on both brand awareness and perceived quality. In both cases we also reject the null hypothesis that advertising plays no

17

lagged brand awareness

advertising

advertising2

POLS 0.942 *** (0.00602) 0.00535 *** (0.00117) -0.00000409 *** (0.000000979) *** *** *** *** ** ** ** **

FE 0.223 *** (0.0479) 0.00687 (0.00443) -0.00000139 (0.00000332)

DGMM 0.679 *** (0.109) 0.0152 (0.0139) -0.0000105 (0.00000745)

SGMM 0.837 *** (0.0266) 0.00627 ** (0.00300) -0.00000524 ** (0.00000239)

marginal eect of advertising at: mean

25th pctl.

50th pctl.

75th pctl.

0.00481 (0.00107) 0.00527 (0.00116) 0.00514 (0.00113) 0.00470 (0.00105) reject *** do not reject do not reject reject reject do not reject *** **

0.00668 (0.00412) 0.00684 (0.00438) 0.00679 (0.00430) 0.00664 (0.00405)

0.0138 (0.0129) 0.0150 (0.0138) 0.0147 (0.0135) 0.0136 (0.00127)

0.00558 (0.00269) 0.00617 (0.00296) 0.00600 (0.00288) 0.00544 (0.00263)

18 0.969 1148 317 0.494 0.940 0.851 1148 317

reject * do not reject do not reject reject ** do not reject

advertising test: 1 = 2 = 0 specication tests: Hansen J dierence-in-Hansen J Arellano & Bond AR(2) Arellano & Bond AR(3) goodness of t measures: R2 -within R2 -between R2 # obs # brands

819 274

1148 317

Table 4: Brand awareness. *, **, and *** indicate a signicance level of 0.10, 0.05, 0.01, respectively. Standard errors in parenthesis.

lagged perceived quality

POLS 0.970 (0.0110) *** *** *** ***

FE 0.391 (0.0611)

DGMM 0.659 (0.204)

SGMM 1.047 (0.0459)

objective quality 0.981 *** (0.0431)

brand awareness **

advertising

advertising2

0.000218 (0.0000952) -0.000000133 (0.000000107)

0.0000822 (0.000198) 0.0000000408 (0.000000162)

-0.0000195 (0.000969) 0.000000108 (0.000000945)

0.0000219 (0.000205) 0.0000000571 (0.000000231)

0.0000649 (0.000944) 0.0000000807 (0.00000308)

brand awareness 0.937 *** (0.0413) 0.00596 *** (0.00165) -0.000298 (0.000256) 0.000000319 (0.000000267)

marginal eect of advertising at: mean ** ** ** **

25th pctl.

50th pctl.

75th pctl.

19
reject do not reject reject do not reject *** ** do not reject do not reject 0.914 1148 317 0.180 0.952 0.909 1148 317 819 274

0.0002 (0.0000819) 0.000215 (0.0000933) 0.000211 (0.00009) 0.0001965 (0.0000793) do not reject

0.0000877 (0.000180) 0.000083 (0.000195) 0.0000844 (0.000191) 0.0000887 (0.000177)

-5.13e-06 (0.000848) -0.0000174 (0.000952) -0.0000139 (0.000922) -2.32e-06 (0.000825)

0.0000295 (0.000176) 0.0000230 (0.000201) 0.0000249 (0.000194) 0.0000310 (0.000170)

0.0000594 (0.000740) 0.0000642 (0.000917) 0.0000623 (0.000847) 0.0000588 (0.000714) do not reject ** ** ***

-0.000256 (0.000222) -0.000292 (0.000251) -0.000282 (0.000242) -0.000248 (0.000215) do not reject ** *** *** reject reject reject do not reject do not reject do not reject reject do not reject reject do not reject reject do not reject

advertising test: 1 = 2 = 0 specication tests: Hansen J dierence-in-Hansen J Arellano & Bond AR(2) Arellano & Bond AR(3) goodness of t measures: R2 -within R2 -between R2 # obs # brands

1148 317

604 178

1148 317

Table 5: Perceived quality. *, **, and *** indicate a signicance level of 0.10, 0.05, 0.01, respectively. Standard errors in parenthesis. SGMM estimates in columns labeled objective quality and brand awareness.

role in determining brand awareness and perceived quality (1 = 2 = 0). Of course, as mentioned above, POLS accounts for neither unobserved heterogeneity nor endogeneity. In the next columns of Tables 4 and 5 we present FE, DGMM, and SGMM estimates that attend to these issues.7 Regardless of the class of estimator we nd a signicant positive eect of advertising expenditures on brand awareness. With the FE estimator we nd that the marginal eect of advertising on awareness at the mean is 0.00668. It is borderline signicant with a p-value of 0.105 and implies an elasticity of 0.00638 (with a standard error of 0.00392). A one-standard-deviation increase of advertising expenditures increase brand awareness by 0.0408 standard deviations (with a standard error of 0.0251). The rate of depreciation of a brands stock of awareness is estimated to be 1-0.223 or 78% per year. The FE estimator identies the eect of advertising expenditures on brand awareness solely from the within-brand across-time variation. The problem with this estimator is that it does not deal with the endogeneity of the lagged dependent variable on the right-hand side of equation (2) and the potential endogeneity of advertising expenditures. We thus turn to the GMM estimators described in Section 3. We focus on the more ecient SGMM estimator. The coecient on the linear term in advertising expenditures is estimated to be 0.00627 (p-value 0.037) and the coecient on the quadratic term is estimated to be 0.00000524 (p-value 0.028). These estimates support the hypothesis that the relationship between advertising and awareness is nonlinear. The marginal eect of advertising on awareness is estimated to be 0.00558 (p-value 0.038) at the mean and implies an elasticity of 0.00533 (with a standard error of 0.00257). A one-standard-deviation increase of advertising expenditures increase brand awareness by 0.0340 standard deviations (with a standard error of 0.0164). The rate of depreciation decreases substantially after correcting for endogeneity and is estimated to be 1-0.828 or 17% per year, thus indicating that an increase in a brands stock of awareness due to an increase in advertising expenditures persists for years to come. The Hansen J test for overidentifying restrictions indicates that the instruments taken together as a group are valid. Recall from Section 3 that we must assume
The estimates use at most 317 out of 348 brands because we restrict the sample to brands with data for two years running but use third and higher lags of brand awareness respectively perceived quality and advertising expenditures as instruments. Dierent sample sizes are reported for the DGMM and SGMM estimators. Sample size is not a well-dened concept in SGMM since this estimator essentially runs on two dierent samples simultaneously. The xtabond2 routine in STATA reports the size of the transformed sample for DGMM and of the untransformed sample for SGMM.
7

20

that an extra condition holds in order for the SGMM estimator to be appropriate. The dierence-in-Hansen J test conrms that it does, as we cannot reject the null hypothesis that the additional instruments for the level equations are valid. While we reject the hypothesis of no second-order serial correlation in the error terms, we cannot reject the hypothesis of no third-order serial correlation. This result further validates our instrumenting strategy. However, one may still be worried about the SGMM estimates because DGMM uses a strict subset of the orthogonality conditions of SGMM and we reject the Hansen J test for the DGMM estimates (see Table 4). From a formal statistical point of view, rejecting the smaller set of orthogonality conditions in DGMM is not conclusive evidence that the larger set of orthogonality conditions in SGMM are invalid (Hayashi 2000, pp. 218221). In Figure 2 we plot the marginal eect of advertising expenditures on brand awareness over the entire range of advertising expenditures for our SGMM estimates along with a histogram of advertising expenditures. For advertising expenditures between $400 million and $800 million per year the marginal eect of advertising on awareness is no longer signicantly dierent from zero and, statistically, it is actually negative for very high advertising expenditures over $800 million per year. The former case covers around 1.9% of observations and the latter less than 0.5%. One possible interpretation is that brands with very high current advertising expenditures are those that are already well-known (perhaps because they have been heavily advertised over the years), so that advertising cannot further boost their awareness. Indeed, average awareness for observations with over $400 million in advertising expenditures is 74.94 as compared to 69.35 for the entire sample. Turning from brand awareness in Table 4 to perceived quality in Table 5, we see that the positive eect of advertising expenditures on perceived quality found by the POLS estimator disappears once unobserved heterogeneity is accounted by the FE, DGMM, and SGMM estimators. In fact, we cannot reject the null hypothesis that advertising plays no role in determining perceived quality. Figure 3 graphically illustrates the absence of an eect of advertising expenditures on perceived quality at the margin for our DGMM estimates. While the eect of advertising expenditures on perceived quality is very imprecisely estimated, it appears to be economically insignicant: The implied elasticity is 0.0000534 (with a standard error of 0.00883) and a one-standard-deviation increase of advertising expenditures decrease perceived quality by 0.000869 standard deviations (with a standard error of 0.144). Note that the comparable eects for brand awareness are two orders of magnitude larger. Much of the remainder of this paper is concerned with demonstrating the robustness of this 21

Marginal effect .004 0 .004 0

200

400

600 800 Advertising expenditures (millions of $) marginal effect of advertising lower 90% confidence limit

1000

1200

1400

upper 90% confidence limit

Density 0 0 .005

.01

.015

200

400

600 800 Advertising expenditures (millions of $)

1000

1200

1400

Figure 2: Pointwise condence interval for the marginal eect of advertising expenditures on brand awareness (upper panel) and histogram of advertising expenditures (lower panel). SGMM estimates. negative result. Before proceeding we note that whenever possible we focus on the more ecient SGMM estimator. Unfortunately, for perceived quality in many cases, including that in the fourth column of Table 5, the dierence-in-Hansen J test rejects the null hypothesis that the extra moments in the SGMM estimator are valid. In these cases we focus on the DGMM estimator.

5.1

Objective and Perceived Quality

An important component of a brands perceived quality is its objective quality. To the extent that objective quality remains constant, it is absorbed into the brand eects. But, even though the time frame of our sample is not very long, it is certainly possible that the objective quality of some brands has changed over the course of our sample. If so, then the lack of an eect of advertising expenditures on perceived quality may be explained if brand managers increase advertising expenditures to compensate for decreases in objective quality. To the extent that increased advertising expenditures 22

Marginal effect .001 0 .001 0

200

400

600 800 Advertising expenditures (millions of $) marginal effect of advertising lower 90% confidence limit

1000

1200

1400

upper 90% confidence limit

Density 0 0 .005

.01

.015

200

400

600 800 Advertising expenditures (millions of $)

1000

1200

1400

Figure 3: Pointwise condence interval for the marginal eect of advertising expenditures on perceived quality (upper panel) and histogram of advertising expenditures (lower panel). DGMM estimates. and decreased objective quality cancel each other out, their net eect on perceived quality may be zero. The diculty with testing this alternative explanation is that we do not have data on objective quality. We therefore exclude from the analysis those categories with brands that are likely to undergo changes in objective quality (appliances, automobiles, computers, consumer electronics, fast food, footwear, pharmaceutical OTC, telecommunications, toys, and travel). The resulting estimates are reported in Table 5 under the heading objective quality. We still nd no eect of advertising expenditures on perceived quality.8

5.2

Variation in Perceived Quality

Another possible reason for the lack of an eect of advertising expenditures on perceived quality is that perceived quality may not vary much over time. This is not
The marginal eects are calculated at the mean, 25th, 50th, and 75th percentile for advertising for the brands in the categories judged to be stable in terms of objective quality over time.
8

23

the case in our data. Indeed, the standard deviation of the year-to-year changes in perceived quality is 0.2154. Even for those products whose objective quality does not change over time there are important changes in perceived quality (standard deviation 0.2130). For example, consider bottled water where we expect little change in objective quality over time, both within and across brands. Nonetheless, there is considerable variation in perceived quality. The perceived quality of Aquana Water range across years from 6.33 to 6.90 and that of Poland Spring Water from 5.91 to 6.43, so the equivalent of over two standard deviations. Across the brands of bottled water the range is from 5.88 to 6.90, or the equivalent of over four standard deviations. Further evidence of variation in perceived quality is provided by the automobiles category. Here we have obtained measures of objective quality from Consumer Reports that rate vehicles based on their performance, comfort, convenience, safety, and fuel economy. We can nd examples of brands whose objective quality does not change at least for a number of years while their perceived quality uctuates considerably. For example, Chevy Silverados objective quality does not change between 2000 and 2002, but its perceived quality increases from 6.08 to 6.71 over these three years. Similarly, GMC Sierras objective quality does not change between 2001 and 2003, but its perceived quality decreases from 6.72 to 6.26. The nal piece of evidence that we have to oer is the variance decomposition from Section 4 (see again Table 3 and Figure 1). Recall that the across-brands standard deviation of brand awareness is about 6 times larger than the within-brand standard deviation. In case of perceived quality the ratio is about 4. Hence, while there is more cross-sectional than time-series variation in our sample, the time-series variation is substantial for both brand awareness and perceived quality. Also recall from Section 4 that perceived quality with an intertemporal correlation of 0.95 is somewhat less persistent than brand awareness with an intertemporal correlation of 0.98. Given that we are able to detect an eect of advertising expenditures on brand awareness, it seems unlikely that insucient variation within brands can explain the lack of an eect of advertising expenditures on perceived quality; instead, our results suggest that the variation in perceived quality is unrelated to advertising expenditures. The question then becomes what besides advertising may drive these changes in perceived quality. There are numerous possibilities, including consumer learning and word-of-mouth eects. Unfortunately, given the data available to us, we cannot further explore these possibilities.

24

5.3

Brand Awareness and Perceived Quality

Another concern is that consumers may confound awareness and preference. That is, consumers may simply prefer more familiar brands over less familiar ones (see Zajonc 1968). To address this issue we proxy for consumers familiarity by adding brand awareness to the regression for perceived quality. The resulting estimates are reported in Table 5 under the heading brand awareness. While there is a signicant positive relationship between brand awareness and perceived quality, there is still no evidence of a signicant positive eect of advertising expenditures on perceived quality.

5.4

Competitive Eects

Advertising takes place in a competitive environment. Most of the industries being studied here are indeed oligopolies, which suggests that strategic considerations may inuence advertising decisions. We next allow a brands stocks of awareness and perceived quality to be aected by the advertising of its competitors as discussed in Section 2.9 Competitors advertising, in turn, can enter our estimation equations (1) and (2) either relative in the share-of-voice specication or absolute in the totaladvertising specication. We report the resulting estimates in Table 6. Somewhat surprisingly, the share-of-voice specication yields an insignicant effect of own advertising. We conclude that the share-of-voice specication is simply not an appropriate functional form in our application. The total-advertising specication readily conrms our main ndings presented above that own advertising aects brand awareness but not perceived quality. This is true even if we allow competitors advertising to enter quadratically in addition to linearly. Competitors advertising has a signicant negative eect on brand awareness and a signicant positive eect on perceived quality. Repeating the analysis using the sum instead of the average of competitors advertising yields largely similar results except that the share-of-voice specication yields a signicant negative eect of advertising on brand awareness, thereby reinforcing our conclusion that this is not an appropriate functional form.10
For this analysis we take the subcategory rather than the category as the relevant competitive environment. Consider for instance the beer, wine, liquor category. There is no reason to expect the advertising expenditures of beer brands to aect the perceived quality or awareness of liquor brands. We drop any subcategory in any year where there is just one brand due to the lack of competitors. 10 We caution the reader against reading too much into these results: The number and identity of the brands within a subcategory or category varies sometimes widely from year to year in the Brandweek Superbrands surveys. Thus, the sum of competitors advertising is an extremely volatile
9

25

lagged awareness/quality

total advertising brand awareness perceived quality 0.845 *** 0.356 ** (0.0217) (0.145)

relative advertising

(relative advertising)2 ** ** *

share of voice brand awareness perceived quality 0.872 *** 1.068 *** (0.0348) (0.0406) 0.236 0.0168 (0.170) (0.0164) -0.00912 -0.00102 (0.0104) (0.00132)

advertising

advertising2

competitors advertising

0.00892 (0.00387) -0.00000602 (0.00000248) -0.00609 (0.00363) ** ** ** **

-0.0000180 (0.000592) -0.0000000303 (0.000000535) 0.00128 (0.000515)

**

marginal eect of advertising at: mean

25th pctl.

26
do not reject do not reject * *** do not reject do not reject reject do not reject 1147 317 ** reject do not reject reject do not reject 1147 317

50th pctl.

75th pctl.

0.00333 ( 0.00239) 0.0164 (0.01218) 0.00624 (0.00448) 0.00264 (0.00190)

0.000225 (0.000218) 0.00113 (0.00110) 0.00429 (0.000416) 0.000179 (0.000173)

0.00812 (0.00355) 0.00881 (0.00382) 0.00861 (0.00375) 0.00797 (0.00349) reject

-0.000140 (0.000524) -0.0000174 (0.000582) -0.0000164 (0.000565) -0.0000132 (0.000510) ** do not reject do not reject ** reject do not reject 1147 317 *** do not reject do not reject reject do not reject 1147 317

advertising test: 1 = 2 = 0 specication tests: Hansen J dierence-in-Hansen J Arellano & Bond AR(2) Arellano & Bond AR(3) # obs # brands

Table 6: Competitive eects. *, **, and *** indicate a signicance level of 0.10, 0.05, 0.01, respectively. Standard errors in parenthesis. DGMM estimates in column labeled total advertising/perceived quality and SGMM estimates otherwise.

Overall, the inclusion of competitors advertising does not seem to inuence our results about the role of own advertising on brand awareness and perceived quality. This justies our focus on the simple model without competition. Moreover, it suggests that the following alternative explanation for our main ndings presented above is unlikely.Suppose awareness depended positively on the total amount of advertising in the brands subcategory or category while perceived quality depended positively on the brands own advertising but negatively on competitors advertising. Then the results from the simple model without competition could be driven by an omitted variables problem: If the brands own advertising is highly correlated with competitors advertising, then we would overstate the impact of advertising on awareness and understate the impact on perceived quality. In fact, we might nd no impact of advertising on perceived quality at all if the brands own advertising and competitors advertising cancel each other out.

5.5

Category-Specic Eects

Perhaps the ideal data for analyzing the eect of advertising are time series of advertising expenditures, brand awareness, and perceived quality for the brands being studied. With long enough time series we could then try to identify for each brand in isolation the eect of advertising expenditures on brand awareness and perceived quality. Since such time series are unfortunately not available, we have focused so far on the aggregate eect of advertising expenditures on brand awareness and perceived quality, i.e., we have constrained the slope parameters in equations (1) and (2) that determine the eect of advertising to be the same across brands. Similarly, we have constrained the carryover parameters in equations (1) and (2) that determine the eect of lagged perceived quality and brand awareness respectively to be the same across brands. As a compromise between the two extremes of brands in isolation versus all brands aggregated, we rst examine the eect of advertising in dierent categories. This adds some cross-sectional variation across the brands within a category. As the rst column of Table 7 shows, for the majority of categories, there is nevertheless insucient variation to identify an eect of advertising even on awareness: There is a signicant positive eect of advertising expenditures on brand awareness for ve categories. At the same time, there is a signicant positive eect on perceived quality for ve
measure of the competitive environment. Moreover, the number of brands varies from 3 for some subcategories to 10 for others, thus making the sum of competitors advertising dicult to compare across subcategories.

27

categories (third column). Two caveats are in order. First, we may be capturing the relationship between advertising expenditures and perceived quality across brands: Because the SGMM estimator adds the equations in levels, it relies on more of the cross-sectional variation for identication. Indeed, the FE estimator that relies solely on variation over time within brands detects a signicant positive eect of advertising expenditures on perceived quality for just two categories. Second, we are pushing the limit on the number of instruments. Indeed, we are unable to obtain estimates unless we collapse the set of instruments, creating one instrument for each variable and lag, rather than one for each period, variable, and lag. Second we examine the carryover rate in dierent categories. As the second column of Table 7 shows, the rate of depreciation for brand awareness ranges from 1-0.875 or 12% for health and beauty to 1-0.751 or 25% for pharmaceutical prescription. For perceived quality the rate of depreciation similarly ranges from 1-0.880 or 12% for appliances to 1-0.766 or 23% for telecommunications (fourth column). Surprisingly, the SGMM estimates indicate that, once we allow for the carryover rate to vary by category, advertising expenditures have a positive signicant eect on perceived quality (0.00106 at the mean), although this is not the case for the FE estimates. In sum, it appears that there are important dierences between categories. These dierences, in turn, may help to explain why some brands advertise heavily despite already enjoying a high level of brand awareness. A case in point is the fast food category. On average, this category exhibits the highest level of brand awareness and the second-highest level of advertising expenditures after telecommunications (see again Table 2). A brand has an incentive to put substantial resources into advertising if it either has a particularly high response to advertising or a particularly high rate of depreciation. For brands in the fast food category the rate of depreciation is 1-0.859 or 14% for brand awareness (compared to 1-0.837 or 16% in the overall sample) and 1-0.849 or 15% for perceived quality. At the same time, however, the marginal eect of advertising expenditures is 0.0144 for brand awareness (compared to 0.00558 in the overall sample) and 0.000727 for perceived quality. Hence, the response to advertising is particularly high for both brand awareness and perceived quality.

5.6

Functional Form

Throughout we consider a quadratic functional form for the eect of the level of advertising expenditures on the level of brand awareness and perceived quality. In

28

Appliances Automobiles Beer, Wine, Liquor Beverages Computers Consumer Electronics Cosmetics and Fragrances Credit Cards Fast Food Food Footwear Health and Beauty Household Petrol Pharmaceutical OTC Pharmaceutical Prescription Telecommunications Toys Travel

brand awareness marginal eect carryover 0.0233 0.838 (0.0167) (0.0730) 0.00526 0.840 (0.0154) (0.0402) -0.0264 0.839 (0.0423) (0.0408) -0.0245 0.869 (0.0554) (0.0265) 0.0193 ** 0.799 (0.00777) (0.0370) 0.0210 ** 0.810 (0.00931) (0.0361) -0.104 * 0.766 (0.557) (0.0521) 0.00983 * 0.834 (0.00527) (0.0371) 0.0144 *** 0.859 (0.00543) (0.0262) 0.0296 0.869 (0.0371) (0.0301) -0.0120 0.830 (0.0248) (0.0622) 0.0841 *** 0.875 (0.0319) (0.0278) 0.0743 0.862 (0.0670) (0.0317) -0.0600 0.847 (0.0676) (0.0357) 0.0147 0.840 (0.206) (0.0604) -0.00683 0.751 (0.0355) (0.0747) 0.0105 0.800 (0.0117) (0.0361) 0.0574 0.815 (0.0673) (0.0761) -0.0982 0.832 (0.104) (0.0415)

rate *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***

perceived marginal eect 0.00374 (0.00315) 0.000172 (0.000864) -0.000988 (0.00535) -0.000564 (0.00567) 0.000722 (0.000470) 0.00189 *** (0.000518) 0.000874 (0.00141) 0.000231 (0.000222) 0.000727 *** (0.000207) 0.00287 (0.00390) 0.00390 *** (0.00139) 0.00665 *** (0.00188) 0.00914 *** (0.00296) 0.00433 (0.00266) 0.00329 (0.00253) -0.00488 ** (0.00199) 0.000203 (0.000497) 0.0000834 (0.00116) 0.00603 (0.00518)

quality carryover 0.880 (0.0413) 0.854 (0.0476) 0.811 (0.0553) 0.877 (0.0463) 0.826 (0.0488) 0.849 (0.0445) 0.862 (0.0545) 0.853 (0.0514) 0.849 (0.0530) 0.873 (0.0432) 0.878 (0.0498) 0.879 (0.0441) 0.876 (0.0434) 0.852 (0.0505) 0.866 (0.0437) 0.800 (0.0521) 0.766 (0.0728) 0.862 (0.0715) 0.861 (0.0465)

rate *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***

Table 7: Category-specic eects. Marginal eect of advertising expenditures on brand awareness and perceived quality at mean by category. Carryover rate by category. *, **, and *** indicate a signicance level of 0.10, 0.05, 0.01, respectively. Standard errors in parenthesis. SGMM estimates.

29

ln(brand awareness)

ln(perceived quality)

lagged awareness/quality 0.825 (0.0281) *** 0.666 (0.201) ***

brand awareness 0.886 *** (0.0459)

perceived quality 0.680 *** (0.226)

lagged ln(awareness)/ln(quality)

ln(advertising) * ** ***

ln(advertising)2

-0.119 (0.775) -0.00137 (0.108)

-0.00557 (0.0596) -0.0337 (0.0190)

advertising

advertising2

0.000136 (0.0000526) -0.000000115 (4.14e-08) ** ** ** ** ** ** ** **

-0.0000143 (0.000146) 3.13e-08 (0.000000152)

marginal eect of advertising at: mean

25th pctl.

30
do not reject *** ** ** do not reject do not reject reject do not reject 795 267 *** reject reject reject do not reject 1123 312

50th pctl.

75th pctl.

-0.00197 (0.00612) -0.0131 (0.0424) -0.00494 (0.0134) -0.00165 (0.00539) reject

-0.00435 (0.00211) -0.0166 (0.00800) -0.00869 (0.00414) -0.00379 (0.00184)

0.000120 (0.0000473) 0.000133 (0.0000518) 0.000130 (0.0000505) 0.000117 (0.0000463)

-0.0000102 (0.000127) -0.0000137 (0.000143) -0.0000127 (0.0000138) -0.00000937 (0.0000123) ** ** do not reject do not reject reject do not reject 819 274 *** reject do not reject do not reject do not reject 1148 317

advertising test: 1 = 2 = 0 specication tests: Hansen J dierence-in-Hansen J Arellano & Bond AR(2) Arellano & Bond AR(3) # obs # brands

Table 8: Functional form. *, **, and *** indicate a signicance level of 0.10, 0.05, 0.01, respectively. Standard errors in parenthesis. SGMM estimates in columns labeled brand awareness and ln(brand awareness) and DGMM estimates in columns labeled perceived quality and ln(perceived quality).

Table 8 we report results for the estimation of equations (1) and (2) using dierent functional forms. We rst consider the case of f (Eit1 ) = 1 ln Eit1 + 2 (ln Eit1 )2 .11 There is no eect of advertising expenditures on brand awareness and we nd a negative eect of advertising on perceived quality. This may indicate that this functional form is not appropriate in our application. Next we allow for a quadratic relationship between the level of advertising expenditures on the one hand and the log of brand awareness and perceived quality on the other hand. As Table 8 shows, we still nd no eect of advertising expenditures on the log of perceived quality. In contrast, we nd a positive eect of advertising on the log of brand awareness.

Discussion

To our knowledge this is the rst study to make use of panel data on a wide range of brands along with recently developed methods for estimating dynamic models to study the eect of advertising on brand awareness and perceived quality. Our panel data allow us to control for the unobserved heterogeneity across brands and to identify the eect of advertising o time-series variation within brands. They also let us account for the endogeneity of advertising. Our main ndings are that advertising expenditures have a signicant positive effect on a brands stock of awareness but no signicant eect on its stock of perceived quality. These ndings are consistent with previous empirical work and laboratory experiments. The results in Ackerberg (2001), for example, indicate that the primary eect of advertising for the particular brand of yogurt being studied is that of informing consumers. However, the importance of information may vary with the stages of a products life cycle. Narayanan & Manchanda (2008) nd that the responsiveness of physicians to the informative content of detailing and the responsiveness to the persuasive content are negatively correlated over time. Mitra & Lynch (1995) show that, especially in mature product categories, advertising has a much stronger eect on the size of the consideration set than on the relative strength of preferences. Our research complements and generalizes existing studies by Shachar & Anand (1998), Ackerberg (2001), Narayanan, Manchanda & Chintagunta (2005), and Narayanan
The number of observations diers slightly across specications because the logarithm of zero is not dened. Our conclusions remain unchanged if we replace ln Ejt1 by ln(c + Ejt1 ), where c > 0 is a constant, in order to be able to use all observations.
11

31

& Manchanda (2008) that apply econometric models to discern the role of advertising for a single brand or industry. The key idea used in these studies to distinguish between informative and persuasive advertising is that informed consumers should not be aected (or as aected) by informative advertising as uninformed consumers whereas the eect of persuasive advertising should be independent of the amount of information that is available to consumers. The diculty lies in identifying the amount of information that is available to consumers. The common approach is to proxy for available information with usage experience: once consumers have used the brand, they must be aware of its existence and should know its characteristics, so informative advertising should not aect them any more. Since usage experience is often not directly observable, this empirical strategy is largely limited to newly introduced brands.12 The current paper contributes to our understanding of the nature of advertising in two ways. By using data on over 300 brands across 19 product categories, we are able to say something more general about the eect of advertising than just for a single brand or industry. In addition, our direct measures of the level of information possessed by consumers and of their quality perceptions allow us to study the channel through which advertising aects consumer choice without making assumptions about the amount of information that is available to consumers on the basis of their purchase behavior. While our main ndings highlight advertising as a means of providing information to consumers, there are important dierences between categories. In some categories at least advertising may also be a means of altering quality perceptions. This conclusion suggest that a long enough time series on advertising expenditures, perceived quality, and brand awareness may prove ideal to identify and quantify the various eects of advertising for a specic category (or even for specic brand) in isolation. At the same time, however, our results hint at the role of the competitive environment that cannot be adequately captured without a broad enough cross section. It is furthermore important to note that our analysis focuses on the short-run relationship between advertising expenditures, brand awareness, and perceived quality. That is, we can only say that advertising has no short-run inuence on perceived quality. Again this is dictated by the data. It is of course still possible that advertising aects perceived quality, but only after a period of time. On the other hand,
Anand & Shachar (2004) pursue a dierent methodology that is not limited to newly introduced brands, although the data requirement may prevent more wide-spread application. Their study of advertising for television shows in the form of previews highlights advertising as a vehicle of matching and information rather than an instrument of persuasion.
12

32

given that we have six years worth of data, the short run in our model is fairly long. It is unclear whether, in practice, the time horizon of rms is much longer than that. Therefore, even if advertising had an eect in the long run, brand managers may not be willing to expend resources today in order to reap benets that are that far in the future. Our ndings may also help to resolve the puzzling fact that advertising has little eect on sales (e.g., Assmus, Farley & Lehmann 1984, Lodish, Abraham, Kalmenson, Livelsberger, Lubetkin, Richardson & Steve 1995). Recall that sales are measured as quantity sold times sales price. Since our results suggest that advertising has no eect on perceived quality, it presumably has no eect on consumers willingness to pay. On the contrary, by making consumers aware of more brands, advertising should be procompetitive and put downward pressure on prices. Hence, if price decreases suciently, then sales remain constant or even decrease in response to advertising even if quantity increases.13 Investigating this question further presents a promising venue for further research.

References
Acemoglu, D. & Robinson, J. (2001), A theory of political transition, American Economic Review 91(4), 938963. Ackerberg, D. (2001), Empirically distinguishing between informative and prestige eects of advertising, Rand Journal of Economics 32(2), 316333. Anand, B. & Shachar, R. (2004), Advertising the matchmaker, Working paper, Harvard University, Cambridge. Arellano, M. & Bond, S. (1991), Some tests of specication for panel data: Monte Carlo evidence and an application to employment equations, Review of Economic Studies 58, 277297. Arellano, M. & Bover, O. (1995), Another look at instrumental variable estimation of error component models, Journal of Econometrics 68, 2952. Assmus, G., Farley, J. & Lehmann, D. (1984), How advertising aects sales: Meta analysis of econometric results, Journal of Marketing Research 21(1), 6574.
A long standing problem in estimating the eect of advertising on sales is the so-called data interval bias (Clarke 1976). The impact of advertising is misestimated to the extent that the ow of sales and the ow of advertising are not properly matched up over the course of a period. In our setting, the dependent variables are a brands stocks of perceived quality and awareness. This may mitigate the data interval bias because these stocks encompass all current and previous ows and thus are more likely to pick up the impact of advertising.
13

33

Bass, F., Bruce, N., Majumdar, S. & Murthi, B. (2007), Wearout eects of different advertising themes: A dynamic Bayesian model of the advertising-sales relationship, Marketing Science 26(2), 179195. Blackburn, J. & Clancy, K. (1982), LITMUS: A new product planning model, in A. Zoltners, ed., Marketing Planning Models, North-Holland, Amsterdam, pp. 4362. Blattberg, R. & Golanty, J. (1978), Tracker: An early test market forecasting and diagnostic model for new product planning, Journal of Marketing Research 15, 192202. Blundell, R. & Bond, S. (1998), Initial conditions and moment restrictions in dynamic panel data models, Journal of Econometrics 87, 115143. Bond, S. (2002), Dynamic panel data models: A guide to micro data methods and practice, Working paper CWP09/02, The Institute for Fiscal Studies, London. Boyer, M. & Moreaux, M. (1999), Strategic underinvestment in informative advertising: The cases of substitutes and complements, Canadian Journal of Economics 32(3), 654672. Clarke, D. (1976), Econometric measurement of the duration of advertising eect on sales, Journal of Marketing Research 13(4), 345357. Doganoglu, T. & Klapper, D. (2006), Goodwill and dynamic advertising strategies, Quantitative Marketing and Economics 4(1), 529. Doraszelski, U. & Markovich, S. (2007), Advertising dynamics and competitive advantage, Rand Journal of Economics 38(3), 557592. Dube, J., Hitsch, G. & Manchanda, P. (2005), An empirical model of advertising dynamics, Quantitative Marketing and Economics 3, 107144. Durlauf, S., Johnson, P. & Temple, J. (2005), Growth econometrics, in P. Aghion & S. Durlauf, eds, Handbook of Econometric Growth, Vol. 1A, North-Holland, Amsterdam, pp. 555677. Erdem, T., Keane, M. & Sun, B. (2008), The impact of advertising on consumer price sensitivity in experience goods markets, Quantitative Marketing and Economics 6(2), 139176. Fershtman, C. & Muller, E. (1993), The benets of being small: Duopolistic competition with market segmentation, Review of Industrial Organization 8, 101111. Hayashi, F. (2000), Econometrics, Princeton University Press, Princeton. Kadiyali, V. (1996), Entry, its deterrence, and its accommodaton: A study of the U.S. photographic lm industry, Rand Journal of Economics 27(3), 452478. 34

Kirmani, A. (1990), The eect of perceived advertising costs on brand perceptions, Journal of Consumer Research 17(2), 160171. Kirmani, A. & Wright, P. (1989), Money talks: Perceived advertising expense and expected product quality, Journal of Consumer Research 16(1), 344353. Little, J. (1975), Brandaid: A marketing mix model, part 1-structure, Operations Research 23, 628655. Lodish, L., Abraham, M., Kalmenson, S., Livelsberger, J., Lubetkin, B., Richardson, B. & Steve, M. (1995), How advertising works: A meta-analysis of 389 real world split cable TV advertising experiments, Journal of Marketing Research 32, 125139. Milgrom, P. & Roberts, J. (1986), Price and advertising signals of product quality, Journal of Political Economy 94(4), 796821. Mitra, A. & Lynch, J. (1995), Toward a reconciliation of market power and information theories of advertising eects on price elasticity, Journal of Consumer Research 21(4), 644649. Moorthy, S. & Hawkins, S. (2005), Advertising repetition and quality perception, Journal of Business Research 58(3), 354360. Moorthy, S. & Zhao, H. (2000), Advertising spending and perceived quality, Marketing Letters 11(3), 221233. Naik, P., Mantrala, M. & Sawyer, A. (1998), Planning media schedules in the presence of dynamic advertising quality, Marketing Science 17(3), 214235. Narayanan, S. & Manchanda, P. (2008), Heterogeneous learning and the targeting of marketing communication for new products, Marketing Science forthcoming. Narayanan, S., Manchanda, P. & Chintagunta, P. (2005), Temporal dierences in the role of marketing communicating for new product categories, Journal of Marketing Research 42, 278291. Nedungadi, P. (1990), Recall and consumer consideration sets: Inuencing choice without altering brand evaluations, Journal of Consumer Research 17(3), 263 276. Nelson, P. (1970), Advertising and consumer behavior, Journal of Political Economy 78(2), 311329. Nelson, P. (1974), Advertising as information, Journal of Political Economy 82(4), 729754. Nerlove, M. & Arrow, K. (1962), Optimal advertising policy under dynamic considerations, Economica 29(114), 129142. 35

Roodman, D. (2007), How to do xtabond2: An introduction to dierence and system GMM in Stata, Working paper no. 103, Center for Global Development, Washington. Shachar, R. & Anand, B. (1998), The eectiveness and targeting of television advertising, Journal of Economics and Management Strategy 7, 363396. Sovinsky Goeree, M. (2008), Limited information and advertising in the US personal computer industry, Econometrica 76(5), 10171074. Stigler, G. (1961), The economics of information, Quarterly Journal of Economics 69(3), 213225. Sutton, J. (1991), Sunk costs and market structure, MIT Press, Cambridge. Tellis, G. & Fornell, C. (1988), Relationship between advertising and product quality over the product life cycle: A contingency theory, Journal of Marketing Research 25(1), 6471. Telser, L. (1964), Advertising and competition, Journal of Political Economy 72(6), 537562. Vidale, M. & Wolfe, H. (1957), An operations research study of sales response to advertising, Operations Research 5(3), 370381. Windmeijer, F. (2005), A nite sample correction for the variance of linear ecient two-step GMM estimators, Journal of Econometrics 126(1), 2551. Wooldridge, J. (2002), Econometric analysis of cross section and panel data, MIT Press, Cambridge. Zajonc, R. (1968), Attitudinal eects of mere exposure, Journal of Personality and Social Psychology 9(2), 127. Zhang, J. & Li, H. (2007), Do high birth rates hamper economic growth?, Review of Economics and Statistics 89(1), 110117.

36